Skip to content
Imaging Evidence Path

Why MRI is especially vulnerable to domain shift

Tonight · ~15 min · read · energy: low · setup: none

Chapter 4 established that domain shift is the rule in medical imaging. MRI is its worst case: the “same” anatomy looks numerically different across many independent axes. This short lesson is why MRI radiomics and MRI foundation models are harder than CT, and what to test.

The anchor: non-physical intensity × many acquisition axes

MRI intensity is not a physical scale (lesson 1), so the same anatomy produces different numbers under different acquisition. That non-comparability multiplies across the axes that define an MRI acquisition. A model trained on one combination is being tested under domain shift when any axis changes.

The axes of MRI domain shift

Expect distribution shift from:

  • field strength (1.5 T vs 3 T vs higher)
  • vendor (Siemens / GE / Philips / …)
  • coil (receive array, coverage)
  • sequence and exact parameters (TR/TE/flip/TI)
  • protocol (resolution, fat-suppression method)
  • reconstruction (parallel imaging, compressed sensing, DL — lesson 6)
  • scanner / site (calibration, environment)

This is why MRI radiomics and MRI foundation models are harder than CT: the “same” anatomy looks numerically different across these axes, and there are more of them than in CT (where HU is at least approximately comparable). TotalSegmentator having a separate MRI model (PMID 39964271) is itself evidence of the shift: you cannot simply run a CT-trained segmenter on MRI.

What to do about it

The discipline is the same as in Chapter 4, applied to more axes:

  • Document the acquisition of training and test data across all these axes.
  • Test the shift explicitly: hold out a vendor, a field strength, a coil type.
  • Normalise/harmonise where defensible (intensity normalisation, bias-field correction, ComBat on features — with its assumptions, Ch. 3), but remember harmonisation adjusts distributions, not physics.
  • Read “generalisable” claims with the axes in mind — “generalisable across 1.5 T Siemens knee MRI” is a scoped claim; “generalisable” with no scope is not (Ch. 7).
Stop and think — then reveal

An MRI radiomics model is trained on 3 T Siemens knee MRIs from one hospital and reports strong internal AUROC. The deployment site is 1.5 T GE. Should you expect the AUROC to hold, and what is the minimum honest test?

No — you should expect a drop, because field strength, vendor, coil, and likely protocol and reconstruction all differ, and MRI intensity/texture are not scanner-comparable (lesson 1). The minimum honest test is frozen-model evaluation on 1.5 T GE data from the deployment site (with a frozen threshold), reporting the drop and recalibrating if needed (Ch. 4). Reporting only the 3 T Siemens internal AUROC as the deployment expectation would be an overclaim of scope.

What to retain

  1. MRI is the worst case for domain shift because intensity is non-physical and there are many acquisition axes (field, vendor, coil, sequence, protocol, reconstruction, site).
  2. “Same” anatomy looks numerically different across these axes — MRI radiomics and foundation models are harder than CT.
  3. Document the axes, test the shift (hold out a vendor/field/coil), harmonise only where defensible, and scope “generalisable” claims to the axes actually tested.

Next: the synthesis question that ties MRI back to Chapter 6 — when synthetic/multimodal MRI asks the model to invent information.